polish: audio cues, Esc cancel, clipboard restore, silence gate, fixed language

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

README.md

@@ -27,5 +27,8 @@ uv run wersper.py --check
 
 ## Notes
 
-- Pasting overwrites your clipboard with the transcription.
-- To change the hotkey or model, edit the constants at the top of `wersper.py`.
+- **Esc** while recording cancels the take.
+- Pop = recording started, Tink = text pasted.
+- Your clipboard is restored after pasting (text only — images are lost).
+- Hotkey, model, language, and silence threshold are constants at the top of
+  `wersper.py`. Set `LANGUAGE = None` for autodetect.

wersper.py

@@ -5,10 +5,12 @@
 """wersper — hold Right Option, speak, release. Text appears at your cursor.
 
 Run:  uv run wersper.py
+Esc while recording cancels the take.
 """
 import subprocess
 import sys
 import threading
+import time
 
 import numpy as np
 import sounddevice as sd
@@ -16,23 +18,35 @@ import mlx_whisper
 from pynput import keyboard
 
 MODEL = "mlx-community/whisper-large-v3-turbo"
+LANGUAGE = "en"  # None = autodetect (slower, occasionally wrong)
 HOTKEY = keyboard.Key.alt_r
 SAMPLE_RATE = 16000
+# ponytail: crude silence gate; raise if background noise gets transcribed,
+# lower if it eats quiet speech
+MIN_RMS = 0.001
 
 frames = []
 stream = None
 kb = keyboard.Controller()
 
 
+def sound(name):
+    subprocess.Popen(["afplay", f"/System/Library/Sounds/{name}.aiff"])
+
+
 def transcribe(audio):
-    return mlx_whisper.transcribe(audio, path_or_hf_repo=MODEL)["text"].strip()
+    result = mlx_whisper.transcribe(audio, path_or_hf_repo=MODEL, language=LANGUAGE)
+    return result["text"].strip()
 
 
 def paste(text):
+    old = subprocess.run("pbpaste", capture_output=True).stdout
     subprocess.run("pbcopy", input=text.encode(), check=True)
     with kb.pressed(keyboard.Key.cmd):
         kb.press("v")
         kb.release("v")
+    time.sleep(0.4)  # let the paste land before giving the clipboard back
+    subprocess.run("pbcopy", input=old)  # ponytail: restores text only, not images
 
 
 def start_recording():
@@ -45,35 +59,46 @@ def start_recording():
         callback=lambda data, *_: frames.append(data.copy()),
     )
     stream.start()
-    print("● recording...", flush=True)
+    sound("Pop")
+    print("● recording... (Esc to cancel)", flush=True)
 
 
-def stop_recording():
+def stop_recording(cancel=False):
     global stream
     if not stream:
         return
     stream.stop()
     stream.close()
     stream = None
+    if cancel:
+        print("✕ cancelled", flush=True)
+        return
     if not frames:
         return
     audio = np.concatenate(frames).flatten()
     if len(audio) < SAMPLE_RATE // 3:  # ignore accidental taps
         return
+    if np.sqrt((audio ** 2).mean()) < MIN_RMS:
+        print("(silence, skipped)", flush=True)
+        return
     # ponytail: transcribe in a thread so the hotkey stays responsive
     threading.Thread(target=lambda: finish(audio), daemon=True).start()
 
 
 def finish(audio):
+    t0 = time.time()
     text = transcribe(audio)
     if text:
-        print(f"→ {text}", flush=True)
+        print(f"→ {text}  ({time.time() - t0:.1f}s)", flush=True)
         paste(text)
+        sound("Tink")
 
 
 def on_press(key):
     if key == HOTKEY:
         start_recording()
+    elif key == keyboard.Key.esc and stream:
+        stop_recording(cancel=True)
 
 
 def on_release(key):
@@ -82,9 +107,9 @@ def on_release(key):
 
 
 def check():
-    """Self-check: model loads and transcribes silence without exploding."""
-    result = transcribe(np.zeros(SAMPLE_RATE, dtype=np.float32))
-    assert isinstance(result, str)
+    """Self-check: model loads, pipeline works, silence gate works."""
+    assert isinstance(transcribe(np.zeros(SAMPLE_RATE, dtype=np.float32)), str)
+    assert np.sqrt((np.zeros(SAMPLE_RATE) ** 2).mean()) < MIN_RMS
     print("ok: model loaded, pipeline works")
 
 
@@ -94,6 +119,6 @@ if __name__ == "__main__":
         sys.exit(0)
     print("loading model (first run downloads ~1.6GB)...", flush=True)
     transcribe(np.zeros(SAMPLE_RATE, dtype=np.float32))  # warm up
-    print(f"ready — hold Right ⌥ to dictate, Ctrl+C to quit", flush=True)
+    print("ready — hold Right ⌥ to dictate, Esc cancels, Ctrl+C quits", flush=True)
     with keyboard.Listener(on_press=on_press, on_release=on_release) as listener:
         listener.join()